[
https://issues.apache.org/jira/browse/CASSANDRA-14160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16441124#comment-16441124
]
Jeff Jirsa commented on CASSANDRA-14160:
----------------------------------------
cc [~cnlwsu] as well (we've seen some meaningful perf improvements with
NEVER_PURGE_TOMBSTONES=true, tagging Chris in case he remembers if the benefit
was in the purge evaluator or getFullyExpiredSSTables() )
> maxPurgeableTimestamp should traverse tables in order of minTimestamp
> ---------------------------------------------------------------------
>
> Key: CASSANDRA-14160
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14160
> Project: Cassandra
> Issue Type: Bug
> Components: Compaction
> Reporter: Josh Snyder
> Assignee: Josh Snyder
> Priority: Major
> Labels: performance
> Fix For: 4.x
>
>
> In maxPurgeableTimestamp, we iterate over the bloom filters of each
> overlapping SSTable. Of the bloom filter hits, we take the SSTable with the
> lowest minTimestamp. If we kept the SSTables in sorted order of minTimestamp,
> then we could short-circuit the operation at the first bloom filter hit,
> reducing cache pressure (or worse, I/O) and CPU time.
> I've written (but not yet benchmarked) [some
> code|https://github.com/hashbrowncipher/cassandra/commit/29859a4a2e617f6775be49448858bc59fdafab44]
> to demonstrate this possibility.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
